The one-screen version: the single most important AI number this week is not a valuation — it's the gap between two lines. Ramp's card-spend data shows token volume still climbing while token spend has barely grown for about two months at the heaviest buyers, because the blended price per token is falling fast (one four-week window: volume +52.4%, spend +23.0%, price −25.4%). The labs are driving that deflation themselves — GPT-6.1 Terra lists at $1/$5 per million tokens, roughly a quarter of Claude Opus 5.5's $4/$20, at near-parity on OpenAI's own charts. And it all lands the week Anthropic walks toward its IPO. Here's what each one asks of you.

1. The demand curve went flat — but read it correctly#

For two years every projection in this industry has been a straight line up and to the right, and the line was dollars. This week that line bent. Per Ramp's closely-watched spend index, spending on OpenAI and Anthropic among the heaviest corporate buyers has grown almost nothing for roughly two months — and because the top 1% of firms drive about 80% of that spend in Ramp's sample, a plateau there is the plateau.

The mistake would be to read "spend stalled" as "demand stalled." It's the opposite. Over one four-week window Ramp clocked token volume up 52.4% against spend up only 23.0% — the difference is a 25.4% drop in the blended price per token. People are using more AI and paying less in aggregate. That is not a bust. It is deflation, and deflation this steep is the defining economic fact of building on models right now.

The usual caveats apply and matter: Ramp is measuring card spend by businesses on its own platform, not the labs' revenue, and a sample is not a census. But the shape is corroborated across weeks, and it rhymes with what we flagged when the AI capex story first wobbled at the start of the month.

What it changes for you: re-forecast on tokens consumed, not dollars spent, and assume your per-token cost keeps falling at a double-digit clip per quarter. If you budgeted AI as a line that only grows, you've mispriced your own gross margin — in your favor. The teams that win the next year are the ones that bank the deflation instead of having already handed it to customers in a discount they can't take back.

2. The labs are undercutting their own flagships#

Why is the blended price collapsing? Because the people who set it want it to. OpenAI's GPT-6.1 Terra lists at roughly $1 per million input / $5 per million output tokens — about a quarter of Claude Opus 5.5's $4 / $20 (current pricing side-by-side here). And on OpenAI's own published charts, Terra at full effort lands near Opus 5.5 on capability (about 60% vs 63% on their task benchmark) at a fraction of the cost per task.

Sit with what that means. The second-best model is now "good enough" for most production work at a quarter of the price of the best one — so the leaders cut their own rates rather than cede the volume. That's not a promotion; it's the structure of a commoditizing market. We argued months ago that the model was never the moat; this is the P&L finally saying it out loud.

What it changes for you: before you renew anything, re-run your evals on the cheap tier — near-parity at a quarter of the price should be your default assumption now, not a surprise. And do not sign a long prepaid token commitment into a falling market; you'd be locking tomorrow's budget at today's price. The flip side is the opportunity: if your competitor is still paying flagship rates for a job the cheap tier does, your unit economics beat theirs on the same product.

3. Anthropic walks into its IPO on exactly this data#

Now layer the capital markets on top. Anthropic — Nasdaq ticker ANT, S-1 filed October 26 — is reportedly lining up an IPO roadshow as early as the week of November 9, targeting a debut before Thanksgiving. It arrives the same month OpenAI's latest round came in around $11B at a ~$1.25T valuation, below its reported $30B / $1.4T target.

So the biggest AI listing ever is about to ask the public market to price token revenue — in the exact week the most-watched spend data says that revenue's growth is flattening while usage soars. The pre-IPO private valuations floated earlier this year (well above a trillion) were underwritten by the straight-line assumption. The amended S-1's price range will be the first real verdict on what that revenue is worth when you price in the deflation, not past it.

What it changes for you: you don't trade the IPO, but you should read its price range like a weather report. Whatever multiple the market puts on "revenue that is mostly resold tokens" is the multiple your investors and acquirers will reach for when they value you. If your business is a thin markup on a model call, that number is your warning. The durable move is the one we keep coming back to: own something the token price can't deflate — proprietary data, workflow lock-in, a distribution channel, a result the customer can't get by calling the API themselves. Price your margin above the model, because the margin inside the model is being competed to zero in public, this quarter, in front of you.

The throughline: this is the week the AI economy stopped being a story about scarcity and started being a story about abundance — cheaper tokens, more usage, thinner margins, a public market about to put a number on all of it. Abundance is wonderful if you sell something scarce on top of it, and brutal if the abundant thing is your product. Figure out which one you are before the S-1 does it for you.